Báo cáo hóa học: " Space-Time Coded OFDM with Low PAPR Anand Venkataraman, Harish Reddy, and Tolga M. Duman" pptx

Hindawi Publishing Corporation
EURASIP Journal on Applied Signal Processing
Volume 2006, Article ID 87125, Pages 1–9
DOI 10.1155/ASP/2006/87125
Space-Time Coded OFDM with Low PAPR
Anand Venkataraman, Harish Reddy, and Tolga M. Duman
Department of Electrical Engineering, Arizona State University, Tempe, AZ 85287-5706, USA
Received 11 January 2005; Revised 25 July 2005; Accepted 1 September 2005
Recommended for Publication by Alex Kot
Recently the use of multiple-input multiple-output (MIMO) or thogonal frequency division multiplexing (OFDM) systems has
been proposed for signaling over frequency-selective fading channels. Although various aspects of these systems have been con-
sidered in the literature, the problem of the inherent high peak-to-average power ratio (PAPR) is not examined. In this paper,
we consider PAPR reduction for MIMO-OFDM systems and propose alternate low-complexity algorithms that can be used in
conjunction with the t rellis shaping method. We show that a PAPR reduction in the order of 4-5 dB can be achieved at the cost
of a slight reduction in the spectral efficiency. Furthermore, we compare the trellis shaping technique with other PAPR reduction
techniques such as tone reservation and partial transmit sequences.
Copyright © 2006 Hindawi Publishing Corporation. All rights reserved.
1. INTRODUCTION
Single-carrier modulation together with equalization and
multicarrier modulation, such as orthogonal frequency di-
vision multiplexing (OFDM), are used to overcome the chal-
lenges posed by dispersive channels. OFDM uses a number
of subcarriers which are orthogonal to each other. Data is
placed on each of the subcar riers and can be recovered at the
receiver by exploiting the orthogonality among the subcarri-
ers.
Recently, in addition to the single-input single-output
OFDM systems, space-time coded OFDM systems have been
receiving significant attention. They were first proposed by
Agrawal et al. in order to achieve data rates of 1.5–3 Mbps
over a bandwidth of 1 MHz [1], and it is shown that space-

time coding can be used to achieve high data rates at low
signal-to-noise ratios (SNRs) over different channels with
different multipath delay profiles. In [2], the authors pro-
pose a space-time code for a Rayleigh flat fading chan-
nel which performs wel l for various wireless local area net-
work (WLAN) applications. In [3], the authors present an
algebraic design framework and propose two approaches
for space-time codes in frequency-selective fading channels,
one of which employs OFDM. In this scheme, a frequency-
selective fading channel is converted into a set of flat block
fading channels. Subsequently, an algebraic framework is
employed to exploit the diversity available in the block fading
channels so as to improve the performance of the system.
Although OFDM has many advantages, it has limitations
including high PAPR and carri er frequency offset sensitivity
[4]. Since the complex baseband OFDM signal is formed by
the superposition of many sinusoids with different frequen-
cies, the instantaneous power of the resulting signal may be
larger than the average power of the OFDM signal exhibit-
ing high peaks. It is important to reduce the PAPR because
the high-power amplifiers (HPAs) in a transmitter need to
have a linear region that is much greater than the average
power, making the HPAs expensive and inefficient. When an
HPA with a linear region slightly greater than the average
power is used, the saturation caused by the large peaks will
induce intermodulation distortion. This distortion increases
the bit error rate (BER) and causes spectral widening, which
results in adjacent channel interference [5]. Moreover, regu-
latory bodies specify a peak envelope power limit for a given
band, which means that modulation schemes such as OFDM

resulting in large peak powers cannot be u sed directly [6].
For some important contributions in PAPR reduction, see
for instance [7–10].
Although many aspects of MIMO-OFDM systems have
been addressed, techniques for reducing the PAPR of the re-
sulting OFDM signal are yet to be developed. In our earlier
work [11, 12], some of the existing single-antenna PAPR re-
duction algorithms are extended to MIMO-OFDM systems.
It is recognized that since the PAPR reduction is achieved
without significantly affecting the error rate of the space-time
codes and since there are no in-band distortion and out-of-
band radiation caused, trellis shaping is a promising tech-
nique for PAPR reduction in MIMO-OFDM systems. In this
paper, our objective is to propose PAPR reduction techniques
2 EURASIP Journal on Applied Signal Processing
suitable for MIMO-OFDM systems. We also propose differ-
ent algorithms of varying complexity levels to be used in con-
junction with trellis shaping for MIMO systems as an alter-
native to the one already being used in the literature, namely,
the Viterbi algorithm. Furthermore, to compare the perfor-
mance of the proposed trellis shaping schemes with other
possible alternatives, we also study several other techniques
via some examples.
The rest of the paper is organized as follows. In Section 2,
trellis shaping for MIMO-OFDM systems is discussed. In
Section 3, we present several algorithms that can be used in
conjunction with trellis shaping. In Section 4,acomprehen-
sive set of examples are reported to demonstrate that signifi-
cant PAPR reduction can be obtained with a slight penalt y in
the spectral efficiency of the MIMO-OFDM system. Finally,

conclusions are provided in Section 5.
2. TRELLIS SHAPING FOR REDUCED PAPR
A complex baseband OFDM signal can be expressed as
x( t)
=
1
√
N
N−1

l=0
X
l
e
j2πlt/T
,(1)
where x(t) is the time domain signal, X
l
is the complex data
symbol on the lth subcarrier, T is the OFDM symbol dura-
tion (excluding the guard interval), and N is the number of
subcarriers. PAPR is defined as the ratio of the peak power to
the average power of the OFDM signal which is given by
PAPR
=
max


x( t)



2
E



x( t)


2

,(2)
where E[
·] is the expected value and E[|x(t)|
2
] is the aver-
age power of x(t). The statistical distribution of the PAPR is
usually characterized by the complementary cumulative dis-
tribution function (CCDF) and is given by
CCDF

PAPR
0

=
Pr

PAPR > PAPR
0


=
1 − F
PAPR

PAPR
0

,
(3)
where F
PAPR
is the cumulative distribution function (CDF)
of the PAPR.
Trellis shaping reduces the PAPR of the transmitted se-
quence by adding a valid convolutionally encoded sequence
found using the Viterbi algorithm to it [13, 14]. In trellis
shaping, we use an (n, k, K) convolutional code C
s
,wheren
is the number of output bits, k is the number of input bits,
and K is the constraint length. Other algorithms including
list Viterbi and stack algorithm can also be used in conjunc-
tion with trellis shaping as will be described later in the pa-
per [11, 12]. The original data bits can be recovered at the
receiver using syndrome former decoding.
PAPR for a MIMO-OFDM signal is defined as the max-
imum of the PAPRs among all parallel transmit antenna
branches. PAPR at the ith transmit antenna is defined as the
ratio of the peak power to the average power of an OFDM
symbol in that branch. It can be expressed as

PAPR
MIMO
= max
1≤i≤n
t
PAPR
i
,(4)
where PAPR
i
= max{|x
i
(t)|
2
}/E[|x
i
(t)|
2
], and n
t
is the num-
ber of transmit antennas. Here, E[
|x
i
(t)|
2
] denotes the aver-
age power of the OFDM symbol from the ith transmit an-
tenna.
2.1. Trellis shaped space-time coded

OFDM with reduced PAPR
The block diagram of a trellis-shaped space-time coded
OFDM system is shown in Figure 1 with n
t
transmit and n
r
receive antennas. The main idea of trellis shaping is to add
an (n, k, K) convolutionally coded sequence to the informa-
tion sequence so that the PAPR of the resulting sequence is
below an acceptable threshold. Let G be the generator matrix
and H the corresponding parity check matrix for the con-
volutional code used. The convolutionally coded sequence
should be removed at the receiver in order to obtain the re-
quired information sequence. Also, the convolutional code
sequence added at the transmitter has to be selected care-
fully. To satisfy these two conditions, the following procedure
is used [13, 14]. The input bit sequence, u, is grouped into
blocks u
j
of size (n − k) and multiplied by (H
T
)
−1
,which
is an (n
− k) × n matrix, resulting in blocks z
j
of length n
bits (the need for this operation will be clear after examin-
ing the decoding process). Thus, redundancy is introduced

at this point and it is given by 1
− (n − k)/n. The output se-
quence of z
j
’s is denoted by Z. A valid codeword of the con-
volutional code has to be selected and added ( modulo 2) to
Z. To accomplish this, a procedure similar to the decoding of
the convolutional codes is followed. The only difference is the
specific metric used (w hich will be described later). Let the
path with the least m etric correspond to the code sequence
Y, which is added (modulo 2) to Z resulting in Z

.Through
syndrome former decoding, we can remove Y at the trans-
mitter. The function of the syndrome former decoding can
be represented mathematically as
z

j

H
T

=
z
j
H
T

y

j
H
T
= u

H
T

−1
H
T
= u,(5)
since y
j
is a valid codeword. The sequence Z

is fed to the
space-time encoder and the outputs are transmitted through
the n
t
antennas. At the receiver, the data is space-time de-
coded, converted to bits represented by

Z, and then passed to
the syndrome former decoder. The output of the decoder is
given by
u, which is an estimate of the information sequence
u.
Foreachbranchinthetrelliswithlabely
j

, we assign a
metric
|f
i
k,d
| when proceeding from the current stage, d − 1,
to the next stage, d. In MIMO-OFDM, the metric at stage d
is the maximum of the metrics amongst the individual trans-
mitting branches and it is given by
max



f
1
k,d


,


f
2
k,d


, ,


f

n
t
k,d



k ∈ S
k
,(6)
where
|f
i
k,d
| corresponds to the metric at the ith transmit an-
tenna, 1
≤ d ≤ N/b
l
is the subblock index, b
l
is the sub-
block length given by n/m,where m is the number of bits
Anand Venkataraman et al. 3
(n −k)bits
u
j
(H
T
)
−1
n bits

z
j
+
Decoding
delay
Trell is
shaper
for C
s
y
j
z

j
STC
encoder
OFDM
modn
OFDM
modn
OFDM
modn
.
.
.
n
t
2
1
OFDM

demodn
OFDM
demodn
.
.
.
n
r
1
STC
decoder

z

j
(H
T
)
u
j
(n −k)bits
Figure 1: Block diagram of trellis shaping for PAPR reduction for multiple antennas using space-time codes.
required to represent 2 complex data symbols, and S
k
=
{
0, 1,2, , NL−1} with L denoting the oversampling factor.
For example, let us consider an (8,1,2) convolutional code. If
16-QAMisusedformodulation,weneed4bitstorepresent
a 16-QAM symbol. Since the number of output bits at each

branch is 8, we can modulate two subcarriers. Therefore, the
subblock length is two. Hence, d w ill have values between 1
and N/b
l
. f
i
k,d
is computed recursively using [13]
f
i
k,d
= f
i
k,(d
−1)
+
db
l
−1

l=(d−1)b
l
X
l
e
j2πlk/LN
,(7)
where the second term on the right-hand side corresponds
to a signal obtained using only the subcarriers (d
− 1)b

l
to
db
l
− 1modulatedbyX
l
.
2.1.1. Computation of the sequence Y
To find the sequence Y which will be added modulo-2 to the
sequence Z, the Viterbi algorithm (together with the met-
ric given in (7)) may be used [13]. In the case of space-time
trellis codes, at the start of each frame (OFDM symbol) the
space-time trellis encoder is assumed to be in state 0. Addi-
tionally, to ensure that the trellis ends in the zero state, trellis
shaping is not done for all the subcarriers. Instead, it is per-
formed only for N
−N
f
of them, where N
f
is the number of
symbols needed to force the space-time trellis to end in the
zero state. The sequence Z

= Z

Y along with N
f
×m zeros
is the input to the space-time encoder for one frame (frame

length is selected to be equal to the number of subcarriers,
N).
Viterbi and list Viterbi algorithms
In the Viterbi algorithm, only one surviving path is stored
for each state at each time instance. Since we need to mini-
mize
|f
i
k,d
|, this process is not optimal. That is, when using
the metric without the absolute values as in (7), it is not pos-
sible to remove the path(s) with a worse partial metric merg-
ing at a certain state while guaranteeing the optimality of the
solution. If we could have a similar equation (to (7)) using
the absolute values instead, we could say that the use of the
Viterbi algorithm would be optimal, however that does not
seem to be feasible. Therefore, there might be a possibility
that the metric deleted at the stage d can have a better met-
ric at stage d + 1, when compared to the metric selected at
stage d andextendedtostaged + 1. As an alternative to the
Viterbi algorithm, to improve the performance of the system,
the list Viterbi algorithm with the same metric as in (7)can
also be used. By storing more than one path at each state,
the list Viterbi algorithm provides alternate paths for search-
ing the best possible sequence resulting in an improved PAPR
reduction. However, this adds to the complexity of the algo-
rithm. For example, if two surviving paths are stored at each
time instance, then the complexity is twice as much as that of
the Viterbi algorithm. Therefore, we propose low complex-
ity approaches such as M-andT-algorithms, or tree search

algorithms such as the Fano algorithm.
Precisely, if we measure the computational complexity
of the algorithm by the number of metrics calculated per
OFDM symbol, the Viterbi algorithm calculates N/b
l
×2
K−1
×
2
k
metrics and the list Viterbi algorithm with the list size L
s
is L
s
times more complex.
Stack algorithm
Sequential decoding algorithms including the stack algo-
rithm [15, 16] can also be employed to find a convolutionally
encoded sequence which results in a better PAPR reduction.
In the stack algorithm, different paths with different depths
are stored based on the value of their corresponding metrics,
that is, top of the stack is the path with the least metric. At
each stage, the path at the top of the stack is replaced with
the 2
k
transitions, where k is the number of inputs to the
convolutional encoder at each time instance, and the stack is
reordered. Only the paths corresponding to the lowest met-
rics are retained in the stack. A metric that can be used with
the stack algorithm is given by

M
s
= max
⎛
⎝


f
1
k,d


2



f
1
d


2

,


f
2
k,d



2



f
2
d


2

, ,


f
n
t
k,d


2



f
n
t
d



2

⎞
⎠
,(8)
where
|f
i
d
|
2
 denotes the average power at the ith transmit
antenna at stage d,1
≤ i ≤ n
t
is the antenna index, |f
i
k,d
|
2
de-
notes the instantaneous power, 1
≤ d ≤ N/b
l
is the subblock
index, b
l
is the subblock length, and k ={0, 1,2, , NL−1}
is the oversampling index. Since paths with larger depths are

consistently replaced with paths of lower depth, the metric
4 EURASIP Journal on Applied Signal Processing
M
s
is not computationally efficient, as confirmed by exten-
sive simulations. Therefore, we suggest the alternate metrics,
M
s
1
and M
s
2
:
M
s
1
=
M
s
√
d
,
M
s
2
=
M
s
d
,

(9)
where d is the depth of the path.
For these metrics, the cost function is also normalized by
the depth of the path (or its square root). Through simula-
tions we have found that, although the alternative metrics are
ad hoc, they improve the PAPR reduction performance of the
trellis shaping algorithm and reduce the amount of necessary
computations for the same PAPR reduction performance. An
illustrative example will be provided in the numerical results
section.
3. OTHER LOW-COMPLEXITY ALGORITHMS
3.1. M-algorithm
Since Viterbi algorithm is relatively more complex, we alter-
natively propose the use of lower-complexity algorithms such
as the M-algorithm [15] which works similar to the Viterbi
algorithm, but it has a smaller number of extended paths at
each interval. The metric used is given by (7). The details of
the algorithm are as follows.
(i) At depth d, consider all 2
k
transitions from each of the
M states, where k is the number of input bits to the
trellis encoder.
(ii) Select the best M paths with the least metric.
(iii) Go to stage d + 1 and repeat the process until the depth
of the trellis is reached.
The value of M determines the resulting computational
complexity which is given by N/b
l
× M × 2

k
. Better PAPR
reduction is obtained by selecting a larger value of M as more
states are included at each processing interval. When M is
equal to the number of states in the trellis encoder, the M-
algorithm becomes the Viterbi algorithm.
3.2. T-algorithm
Another algorithm that c an be used to reduce the computa-
tional complexity compared to the Viterbi algorithm is the
T-algorithm [15] which also works similar to the Viterbi al-
gorithm but maintains a variable number of paths based on
a threshold, T. For the T-algorithm, the same metric given in
(7) is used, and the number of surviving states at each inter-
val is determined by the closeness of a path with that of the
best path. The algorithm is described as follows.
(i) At depth d, consider all 2
k
transitions from each of the
surviving states.
(ii) Let the path with the best metric be α.
(iii) Subtract α from each of the metrics and if the differ-
ence is less than a predefined threshold, T, accept the
transition and go to stage d+1. Repeat the process until
the depth of the trellis is reached.
The computational complexity of the algorithm depends
on T and it can be studied through simulations. In general,
the larger the value of T, the higher the computational com-
plexity and the better the resulting PAPR reduction.
3.3. Fano algorithm
The Fano algorithm [15] can also be used to select the se-

quence with reduced PAPR through the trellis such that the
PAPR of the transmitted sequence is less than a predefined
threshold. The algorithm traverses depth first through the
trellis and when the metric becomes larger than a predefined
threshold at a particular stage, the algorithm backtracks to
find an untried path in the preceding stages and proceeds
depth first again.
The algorithm calculates metrics of the 2
k
branches at
stage d. M
s
is used as the metric and the path with the small-
est metric is found. If this metric is lower than the threshold,
we accept the transition and go to stage d + 1. Otherwise, the
algorithm backtracks to stage d
−1 and finds an untried path
with the least metric. If this metric is lower than the thresh-
old, we proceed depth first again through the trellis. If not, we
backtrack to stage d
− 2 and repeat the same process. While
backtracking, if the root node is reached, that is, we cannot
track back any further, we increase the threshold and proceed
depth first all over again. Since we do not have the problem of
paths of higher depth being replaced by paths of lower depth,
that is, the transitions take place only between neighboring
stages, we can use M
s
as the metric. The value of the thresh-
old is determined through simulations in order to optimize

the PA PR reduction and minimize the computational com-
plexity. In general, the lower the threshold, the greater the
computational complexity; however, the better the PAPR re-
duction.
Comparison of algorithms used in
conjunction with trellis shaping
The efficiency of the trellis shaping algor ithms is calculated
in terms of the achieved PAPR reduction and the number of
metrics calculated per OFDM symbol. The Viterbi algorithm
should perform better than the M-andT-algorithms, be-
cause at each processing interval transitions from all of the
states of the trellis encoder are considered, whereas in M-
and T-algorithms, depending on the value of M and T,tran-
sitions from only a few states are considered. By increasing
the value of M and T, the performance of the M-andT-
algorithms should improve, as we include more states at each
processing interval in the trellis.
The stack algorithm is expected to perform better than
the Viterbi, M-, and T-algorithms because the probability
of eliminating a good path decreases [11, 12]. Compared
to the stack algorithm, the Fano algorithm has a smaller
memory requirement. For an appropriate value of the thresh-
old, the Fano algorithm may perform better than the M-, T-
, and Viterbi algorithms because it can find alternate paths
through the trellis. When we compare the Fano and stack al-
gorithms that use the same metric, M
s
, and a proper choice
Anand Venkataraman et al. 5
02468101214

PAPR (dB)
10
−5
10
−4
10
−3
10
−2
10
−1
10
CCDF
Viterbi, 1 antenna
No TS, 1 antenna
Viterbi, Alamouti
No TS, Alamouti
Figure 2: Comparison of CCDFs of PAPR for the single-antenna
case and the two-antenna case with Alamouti scheme. (TS refers to
trellis shaping)
of the threshold for a similar PAPR reduction, the Fano al-
gorithm will potentially calculate a smaller number of met-
rics because the transition takes place only between neigh-
boring nodes. The threshold values for the T-algorithm and
the Fano algorithm are selected based on a trade-off between
computational complexity and PAPR reduction, and they are
found based on simulations.
4. EXAMPLES
In this section, we present results of the PAPR reduction
achieved for space-time coded OFDM system employing the

proposed algorithms for use with trellis shaping. The com-
parison of the PAPR reduction achieved for single- and two-
antenna cases for 128 subcarriers using an (8,1,2) convolu-
tional code, and 16-QAM is given in Figure 2. For the case
with two transmit antennas, the Alamouti scheme [17]isem-
ployed. We observe that the PAPR reduction obtained using
the Alamouti scheme is better than the single antenna case.
For the rest of the examples, we consider a space-time
coded OFDM system with N
= 128, two transmit anten-
nas, and one receive antenna. To compare the performance of
the various algorithms used in conjunction with trellis shap-
ing, we consider an (8,1,4) (8-state), a (4,1,4) (8-state) and
an (8,1,2) (2-state) shaping code. In simulations, we use an
oversampling factor of 4 which is sufficiently accurate for the
discrete samples to model the continuous time signal.
The CCDF of the PAPR for list Viterbi algorithm employ-
ing the Alamouti scheme, (8,1,4) convolutional code, and
4-PSK modulation is given in Figure 3.Itcanbeobserved
that list Viterbi decoding (with list size 4) performs better
(by approximately 0.5 dB) than the Viterbi algorithm. The
CCDF of the resulting PAPR for the Alamouti scheme with
a 16-QAM constellation using the (8,1,4) (8-state) is shown
02468101214
PAPR (dB)
10
−5
10
−4
10

−3
10
−2
10
−1
10
CCDF
No TS
List size
= 4(1.75 b/s/Hz )
Viterbi (1.75 b/s/Hz )
Figure 3: Comparison of CCDFs of PAPR for Viterbi decoding and
list Viterbi decoding with the Alamouti scheme.
0 24681012
PAPR (dB)
10
−5
10
−4
10
−3
10
−2
10
−1
10
CCDF
Viterbi
No TS
M

= 2
M
= 4
T
= 1
T
= 2
Figure 4: Comparison of the CCDFs of the PAPR between Viterbi,
M-, and T-algorithms for the Alamouti scheme with an (8, 1, 4)
shaping code.
in Figure 4. T he original Alamouti scheme has a spectral
efficiency of 4 b/s/Hz and the trellis-shaped Alamouti scheme
has a spectr al efficiency of 3.5 b/s/Hz with a subblock length
of two. The Viterbi algorithm achieves a PAPR reduction of
about 5 dB compared to the uncoded system at a CCDF level
of 10
−4
. At the same CCDF level, compared to the Viterbi al-
gorithm, M-algorithm with M
= 4andM = 2 is inferior by
about 1.7dB and 2.5 dB, respectively, in terms of the PAPR
6 EURASIP Journal on Applied Signal Processing
0 24681012
PAPR (dB)
10
−5
10
−4
10
−3

10
−2
10
−1
10
CCDF
Viterbi
Stack M
s
Fano, threshold=7dB
No TS
Figure 5: Comparison of the CCDFs of the PAPR between Viterbi,
Stack, and Fano algorithms for the Alamouti scheme with an
(8, 1, 2) shaping code.
reduction achieved. PAPR reduction achieved by an 8-state
shaping code using T-algorithm with T
= 1andT = 2 is in-
ferior to the Viterbi algorithm by 3.7and2.5 dB, respectively.
The computational complexity of the Viterbi algorithm,
M-algorithm with M
= 4andM = 2, and T-algorithm with
T
= 1andT = 2 is 1024, 512, 256, 172, and 268, respec-
tively. It can be seen that as the computational complexity is
reduced there is degradation in the PAPR reduction.
The CCDF of the resulting PAPR for the Alamouti
scheme with a 16-QAM constellation using the (8,1,2) (2-
state) is shown in Figure 5. For a 2-state shaping code using
the stack algorithm with the metric M
s

, we obtain a PAPR re-
duction of about 5 dB at a CCDF level of 10
−4
. With the Fano
algorithm, we obtain a reduction in PAPR similar to that of
the stack algorithm using met ric M
s
at a CCDF of 10
−4
.At
the same CCDF level, Viterbi algorithm achieves a PAPR re-
duction of about 4.5 dB. The computational complexities for
the Viterbi, stack, and Fano algorithms with the specific pa-
rameters selected are 256, 550, and 225 per OFDM symbol,
respectively.
The CCDF of the PAPR using the alternate metrics for
stack algorithm using an (8,1,2) convolutional code is given
in Figure 6. We see that using these alternate metrics the loss
in PAPR reduction is within 1 dB. On the other hand, the
computational complexity using M
s
, M
s
1
,andM
s
2
are 550,
210, and 155 per OFDM symbol, respectively. Therefore, us-
ing M

s
1
reduces the computational complexity by half when
compared to M
s
. However, the reduction in PAPR is de-
graded by only around 0.5 dB at CCDF level of 10
−3
.Stack
algorithm with M
s
1
performs similar to the Viterbi algorithm
and the compuational complexities are comparable. By us-
ing M
s
2
, we achieve further reduction in computational com-
plexity but with trade-off in PAPR reduction as also apparent
in the figure.
02468101214
PAPR (dB)
10
−5
10
−4
10
−3
10
−2

10
−1
10
CCDF of the PAPR
No TS
M
s2
M
s1
M
s
Figure 6: Comparison of the CCDFs of the PAPR for alternate met-
rics used with stack algorithm.
The computational complexity and the PAPR at a CCDF
level of 10
−3
for the different algorithms with (8,1,2) and
(8,1,4) shaping codes are summarized in Ta ble 1. From the
simulations it is noted that, when the Fano algorithm is
employed for a 2-state shaping code, the number of met-
rics computed per OFDM symbol is 225 for a threshold of
7 dB (averaged over a large number of simulations) and 170
for a threshold of 7.5 dB. Clearly, using alternate metrics for
stack algorithm reduces the computational complexity. Thus,
atrade-off exists between the selected threshold and the com-
putational complexity.
We now consider the (4,1,4) (8-state) shaping code. The
CCDFs of the resulting PAPRs for the Alamouti scheme with
a 4-PSK constellation using M-, T-, and Fano algorithms are
shown in Figure 7. The spectral efficiency of the uncoded sys-

tem is 2 b/s/Hz. The spectral efficiency of the trellis-shaped
space-time coded OFDM system is 1.5 b/s/Hz. The subblock
length is two. We see from the plots that the reduction in
PAPR using M-algorithm with four states is very close to
that of the Viterbi algorithm. This may be because of the
increase in the redundancy of the convolutional code. At a
CCDF level of 10
−3
, the PAPR is 6.4 dB for the Viterbi al-
gorithm. At the same CCDF level, the PAPRs for the M-
algorithm with M
= 4, T-algorithm with T = 1, and Fano
algorithm with a threshold of 7 dB are 6.6, 8, and 7 dB respec-
tively. The computational complexities for the Viterbi algo-
rithm, M-algorithm with M
= 4, T-algorithm with T = 1,
and Fano algorithm with a threshold of 7 dB are 1024, 512,
232, and 270 per OFDM signal, respectively.
In order to illustrate the performance obtained with
space-time trellis codes, the CCDF of the resulting PAPRs for
the 4-state space-time trellis code from [18] with a (4,1,4)
(8-state) shaping code using M-, T-, and Fano algorithms
are shown in Figure 8. The subblock length is two. QPSK
Anand Venkataraman et al. 7
Table 1: Comparison of the computational complexity and the PAPR at a CCDF of 10
−3
for the different algorithms used in conjunction
with trellis shaping with subblock length (b
l
)=2andN=128 for the Alamouti scheme.

Trellis shaping algorithms Number of metrics computed PAPR at CCDF = 10
−3
(dB)
Viterbi (8,1,2) 256 7.4
Stack M
s
(8,1,2) 550 6.8
Stack M
s1
(8,1,2) 210 7.4
Stack M
s2
(8,1,2) 155 7.8
Fano Threshold
= 7 dB (8,1,2) 225 6.9
Viterbi (8,1,4) 1024 6.6
M
= 2 (8,1,4) 256 9.0
M
= 4 (8,1,4) 512 8.0
T
= 1 (8,1,4) 172 9.7
0 24681012
PAPR (dB)
10
−5
10
−4
10
−3

10
−2
10
−1
10
CCDF
Viterbi
M
= 4
Fano, threshold
= 7dB
T
= 1
No TS
Figure 7: Comparison of the CCDFs of the PAPR for the Alamouti
scheme using a (4, 1, 4) shaping code.
constellation has spectral efficiency of 2 b/s/Hz and hence,
the trellis-shaped space-time coded OFDM system has a
spectral efficiency of 1.5 b/s/Hz. We observe that the reduc-
tion in the PAPR using the M-algorithm wi th four states is
very close to the one achieved by the Viterbi algorithm. At a
CCDF level of 10
−3
, the PAPRs for the Viterbi algorithm, M-
algorithm with M
= 4, T-algorithm with T = 1, and Fano al-
gorithm with a threshold of 7 dB are 7.2, 7.4, 8, and 7 dB, and
the resulting computational complexities are 1024, 512, 244,
and 296 per OFDM symbol, respectively. Hence, low com-
plexity algorithms can be used instead of Viterbi and stack

algorithms for reduction in computational complexity with-
out degrading the performance significantly.
In order to illustrate the effectiveness of the trellis shap-
ing for MIMO OFDM systems, we also study the use of
several other techniques, namely, the use of partial trans-
mit sequences [20] and tone reservation [21]. The com-
parison of the CCDF of the PAPRs obtained using trellis
0 2 4 6 8 10 12
PAPR (dB)
10
−5
10
−4
10
−3
10
−2
10
−1
10
CCDF
M = 4
T
= 1
Viterbi
Fano, threshold
= 7dB
No TS
Figure 8: Comparison of the CCDFs of the PAPR for the space-time
Tre llis code from [18] using a (4, 1, 4) shaping code.

shaping, tone reser vation, and partial transmit sequences
for N
= 128 using the (8,1,4) code and space-time codes
from [17, 19] are shown in Figures 9 and 10,respectively.
Here, we use a 4-PSK constellation which results in a spectral
efficiency of 2 b /s/Hz. All three PAPR reduction techniques
result in a spectral efficiency of 1.75 b/s/Hz. As can be seen
from the figures, trellis shaping performs comparable or bet-
ter than partial transmit sequences and tone reservation in
terms of PAPR reduction, while tone reservation performs
better than the partial transmit sequences.
The bit error rate of the three PAPR reduction techniques
under a quasi-static flat Rayleigh fading channel, which is
constant during the transmission of an OFDM symbol and
changes independently from one symbol to another, is given
in Figure 11. In trellis shaping , the degradation in the BER
is due to the error in the syndrome former decoding and is
the same irrespective of the algorithm used. In partial trans-
mit sequences, if one of the rotational fa ctors is decoded
8 EURASIP Journal on Applied Signal Processing
0 2 4 6 8 10 12 14
PAPR (dB)
10
−5
10
−4
10
−3
10
−2

10
−1
10
CCDF
TS + STC (1.75 b/s/Hz)
PTS + STC (1.75 b/s/Hz)
TR + STC (1.75 b/s/Hz)
STC (2 b/s/Hz)
Figure 9: Comparison of the CCDFs of the PAPR for the STC be-
tween trellis shaping, partial transmit sequence, and tone reserva-
tion for STTC from [19] with N
= 128. (TR:tone reservation, PTS:
partial transmit sequence).
02468101214
PAPR (dB)
10
−5
10
−4
10
−3
10
−2
10
−1
10
CCDF
Alamouti (2 b/s/Hz)
TS + Alamouti (1.75 b/s/Hz)
PTS + Alamouti (1.75 b/s/Hz)

TR + Alamouti (1.75 b/s/Hz)
Figure 10: Comparison of the CCDFs of the PAPR between trel-
lis shaping, partial transmit sequence and tone reservation for the
Alamouti scheme with N
= 128.
incorrectly, then the entire subblock is decoded erroneously,
and this results in the BER degradation. There is no BER
degradation using tone reservation, as no side information
is transmitted to decode the data at the receiver.
In partial transmit sequences, each iteration involves the
rotation of the V subblocks (N subcarriers are divided into V
subblocks) and the computation of an IFFT of size NL where
L is the oversampling factor. Hence, the complexity for each
iteration is given by N rotations, NLlog NL additions, and
NLlog NL multiplications per IFFT. In tone reservation, the
0 5 10 15 20 25
SNR (dB)
10
−3
10
−2
10
−1
10
Probability of bit error
Alamouti (2 b/s/Hz)
TS + Alamouti (1.75 b/s/Hz)
PTS + Alamouti (1.75 b/s/Hz)
TR + Alamouti (1.75 b/s/Hz)
Figure 11: Comparison of the BER for three PAPR reduction tech-

niques with N
= 128.
complexity for each iteration is given by the NL comparisons
to locate the peak, NLmultiplications to scale the signal δ(t),
and NL additions/subtractions for reducing the peak at the
given location.
5. CONCLUSIONS
In this paper, we have considered the problem of PAPR re-
duction for MIMO-OFDM systems. We have extended the
use of trellis shaping to MIMO-OFDM systems using space-
time trellis and space-time block codes. In addition to the
commonly used Viterbi algorithm in the trellis shaping, we
have proposed the use of several other algorithms that pro-
vide lower complexity solutions and/or improved PAPR re-
duction performance. We have observed that with a slight re-
duction in the spectral efficiency of the system, it is possible
to achieve a PAPR reduction in the order of 4-5 dB. We have
also compared the performance of trellis shaping against
tone reservation and partial transmit sequences which are
alternative low-complexity approaches. Our proposed low-
complexity algorithms provide a computational complexity
andPAPRreductionperformancetrade-off.
ACKNOWLEDGMENTS
This work was supported in part by NSF CAREER Award
CCR-9984237 and by a grant from the Connection One Cen-
ter. Also, par t of this work was performed while the third au-
thor was on a sabbatical leave at Bilkent University, Turkey.
REFERENCES
[1]D.Agrawal,V.Tarokh,A.Naguib,andN.Seshadri,“Space-
time coded OFDM for high data-rate wireless communication

Anand Venkataraman et al. 9
over wideband channels,” in Proc. 48th IEEE Vehicular Tech-
nology Conference (VTC ’98), vol. 3, pp. 2232–2236, Ottawa,
Ontario, Canada, May 1998.
[2] S. Rouquette-Leveil and K. Gosse, “Space-time coding options
for OFDM-based WLANs,” in Proc. 55th IEEE Vehicular Tech-
nology Conference (VTC ’02), vol. 2, pp. 904–908, Birming-
ham, Ala, USA, May 2002.
[3] H. El Gamal, A. R. Hammons Jr., Y. Liu, M. P. Fitz, and O. Y.
Takeshita, “On the design of space-time and space-frequency
codes for MIMO frequency-selective fading channels,” IEEE
Transactions on Information Theory, vol. 49, no. 9, pp. 2277–
2292, 2003.
[4] T. Pollet, M. V. Bladel, and M. Moeneclaey, “BER sensitivity of
OFDM systems to carrier frequency offset and Wiener phase
noise,” IEEE Transactions on Communications, vol. 43, no. 234,
pp. 191–193, 1995.
[5] R. O’Neill and L. B. Lopes, “Envelope variations and spectral
splatter in clipped multicarrier signals,” in Proceedings of 6th
IEEE International Symposium on Personal, Indoor and Mo-
bile Radio Communications (PIMRC ’95), vol. 1, pp. 71–75,
Toronto, Ontario, Canada, September 1995.
[6] V. Tarokh and H. Jafarkhani, “On the computation and reduc-
tion of the p eak-to-average power ratio in multicarrier com-
munications,” IEEE Transactions on Communications, vol. 48,
no. 1, pp. 37–44, 2000.
[7] H. Ochiai and H. Imai, “Performance analysis of deliberately
clipped OFDM signals,” IEEE Transactions on Communica-
tions, vol. 50, no. 1, pp. 89–101, 2002.
[8] X. Huang, J. Lu, J. Zheng, K. B. Letaief, and J. Gu, “Compand-

ing transform for reduction in peak-to-average power ratio
of OFDM signals,” IEEE Transactions on Wireless Communi-
cations, vol. 3, no. 6, pp. 2030–2039, 2004.
[9] R. van Nee and A. de Wild, “Reducing the peak-to-average
ratio of OFDM,” in Proceedings of 48th IEEE Vehicular Tech-
nology Conference (VTC ’98), vol. 3, pp. 2072–2076, Ottawa,
Ontario, Canada, May 1998.
[10] X. Wang, T. T. Tjhung, and C. S. Ng, “Reduction of peak-
to-average power ratio of OFDM system using a companding
technique,” IEEE Transactions on Broadcasting, vol. 45, no. 3,
pp. 303–307, 1999.
[11] H. Reddy, “Peak to average power reduction for coded MIMO-
OFDM,” Master’s thesis, Arizona State University, Tempe,
Ariz, USA, December 2003.
[12] H. Reddy and T. M. Duman, “Space-time coded OFDM with
low PAPR,” in Proceedings of IEEE Global Telecommunications
Conference (GLOBECOM ’03), vol. 2, pp. 799–803, San Fran-
cisco, Calif, USA, December 2003.
[13] W. Henkel and B. Wagner, “Another application for trellis
shaping: PAR reduction for DMT (OFDM),” IEEE Transac-
tions on Communications, vol. 48, no. 9, pp. 1471–1476, 2000.
[14] H. Ochiai, “A novel t rellis-shaping design with both peak and
average power reduction for OFDM systems,” IEEE Transac-
tions on Communications, vol. 52, no. 11, pp. 1916–1926, 2004.
[15] C. Schlegel, Trellis Coding, IEEE Press, Piscataway, NJ, USA,
1997.
[16] A. J. Viterbi and J. K. Omura, Principles of Digital Communi-
cation and Coding, McGraw-Hill, New York, NY, USA, 1979.
[17] S. M. Alamouti, “A simple transmit diversity technique for
wireless communications,” IEEE Journal on Selected Areas in

Communications, vol. 16, no. 8, pp. 1451–1458, 1998.
[18] Q. Yang and R. S. Blum, “Optimum space time convolutional
codes,” in Proceedings of IEEE Wireless Communications and
Networking Conference (WCNC ’00), vol. 3, pp. 1351–1355,
Chicago, Ill, USA, September 2000.
[19] V. Tarokh, N. Seshadri, and A. R. Calderbank, “Space-time
codes for high data rate wireless communication: performance
criterion and code construction,” IEEE Transactions on Infor-
mation Theory, vol. 44, no. 2, pp. 744–765, 1998.
[20] S. M
¨
uller, R. B
¨
auml, R. Fischer, and J. Huber, “OFDM with
reduced peak-to-average power ratio by multiple signal repre-
sentation,” Annals of Telecommunications,vol.52,no.1-2,pp.
58–67, 1997.
[21] J. Tellado, Multicarrier Modulation with Low PAR : Applications
to DSL and Wireless, vol. 587 of Kluwer International Series in
Engineering and Computer Science, Kluwer Academic, Boston,
Mass, USA, 2000.
Anand Venkataraman received the B.E.
(honors) degree in electrical engineering
from Birla Institute of Technology and Sci-
ence, Pilani, India, in 2002, and the M.S.
degree in electrical engineering from Ari-
zona State University in 2004. He is cur-
rently with Qualcomm Inc., San Diego. His
research interests include multicarrier com-
munications and spread-spectrum commu-

nications.
Harish Reddy received his B.E. degree from
BMS College of Engineering, Bangalore, In-
dia, in 2000, and the M.S. in electrical en-
gineering from Arizona State University in
2003. He is working in the Wireless Re-
search Group of Tata Consultancy Services
since May 2004. His research interests in-
clude OFDM, signal processing for w ireless
communications, and MIMO systems.
Tolga M. Duman received the B.S. degree
from Bilkent University in 1993, and the
M.S. and Ph.D. degrees from Northeastern
University, Boston, in 1995 and 1998, re-
spectively, all in electrical engineering. Since
August 1998, he has been with the Elec-
trical Engineering Department of Arizona
State University first as an Assistant Profes-
sor (1998–2004), and currently as an Asso-
ciate Professor. His current research inter-
ests are in digital communications, wireless and mobile commu-
nications, channel coding, turbo codes, coding for recording chan-
nels, and coding for wireless communications. He is the recipient of
the National Science Foundation CAREER Award, IEEE Third Mil-
lennium Medal, and IEEE Benelux Joint Chapter Best Paper Award
(1999). He is a Senior Member of IEEE, and an Editor for IEEE
Transactions on Wireless Communications.